From owner-svn-src-stable@freebsd.org Tue Dec 25 07:39:36 2018 Return-Path: Delivered-To: svn-src-stable@mailman.ysv.freebsd.org Received: from mx1.freebsd.org (mx1.freebsd.org [IPv6:2610:1c1:1:606c::19:1]) by mailman.ysv.freebsd.org (Postfix) with ESMTP id EEC4F1345092; Tue, 25 Dec 2018 07:39:35 +0000 (UTC) (envelope-from arybchik@FreeBSD.org) Received: from mxrelay.nyi.freebsd.org (mxrelay.nyi.freebsd.org [IPv6:2610:1c1:1:606c::19:3]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits) server-signature RSA-PSS (4096 bits) client-signature RSA-PSS (4096 bits) client-digest SHA256) (Client CN "mxrelay.nyi.freebsd.org", Issuer "Let's Encrypt Authority X3" (verified OK)) by mx1.freebsd.org (Postfix) with ESMTPS id 933AF8F58F; Tue, 25 Dec 2018 07:39:35 +0000 (UTC) (envelope-from arybchik@FreeBSD.org) Received: from repo.freebsd.org (repo.freebsd.org [IPv6:2610:1c1:1:6068::e6a:0]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (Client did not present a certificate) by mxrelay.nyi.freebsd.org (Postfix) with ESMTPS id 8407390D7; Tue, 25 Dec 2018 07:39:35 +0000 (UTC) (envelope-from arybchik@FreeBSD.org) Received: from repo.freebsd.org ([127.0.1.37]) by repo.freebsd.org (8.15.2/8.15.2) with ESMTP id wBP7dZsw052874; Tue, 25 Dec 2018 07:39:35 GMT (envelope-from arybchik@FreeBSD.org) Received: (from arybchik@localhost) by repo.freebsd.org (8.15.2/8.15.2/Submit) id wBP7dY7L052869; Tue, 25 Dec 2018 07:39:34 GMT (envelope-from arybchik@FreeBSD.org) Message-Id: <201812250739.wBP7dY7L052869@repo.freebsd.org> X-Authentication-Warning: repo.freebsd.org: arybchik set sender to arybchik@FreeBSD.org using -f From: Andrew Rybchenko Date: Tue, 25 Dec 2018 07:39:34 +0000 (UTC) To: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-stable@freebsd.org, svn-src-stable-11@freebsd.org Subject: svn commit: r342455 - stable/11/sys/dev/sfxge X-SVN-Group: stable-11 X-SVN-Commit-Author: arybchik X-SVN-Commit-Paths: stable/11/sys/dev/sfxge X-SVN-Commit-Revision: 342455 X-SVN-Commit-Repository: base MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit X-Rspamd-Queue-Id: 933AF8F58F X-Spamd-Bar: -- Authentication-Results: mx1.freebsd.org X-Spamd-Result: default: False [-2.97 / 15.00]; local_wl_from(0.00)[FreeBSD.org]; NEURAL_HAM_MEDIUM(-1.00)[-0.998,0]; NEURAL_HAM_SHORT(-0.98)[-0.975,0]; NEURAL_HAM_LONG(-1.00)[-0.998,0]; ASN(0.00)[asn:11403, ipnet:2610:1c1:1::/48, country:US] X-BeenThere: svn-src-stable@freebsd.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: SVN commit messages for all the -stable branches of the src tree List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Tue, 25 Dec 2018 07:39:36 -0000 Author: arybchik Date: Tue Dec 25 07:39:34 2018 New Revision: 342455 URL: https://svnweb.freebsd.org/changeset/base/342455 Log: MFC r341785 sfxge(4): use n Tx queues instead of n + 2 on EF10 HW On EF10 HW we can avoid sending packets without checksum offload or with IP-only checksum offload to dedicated queues. Instead, we can use option descriptors to change offload policy on any queue during runtime. Thus, we don't need to create two dedicated queues. Submitted by: Ivan Malov Sponsored by: Solarflare Communications, Inc. Differential Revision: https://reviews.freebsd.org/D18390 Modified: stable/11/sys/dev/sfxge/sfxge.c stable/11/sys/dev/sfxge/sfxge.h stable/11/sys/dev/sfxge/sfxge_ev.c stable/11/sys/dev/sfxge/sfxge_tx.c stable/11/sys/dev/sfxge/sfxge_tx.h Directory Properties: stable/11/ (props changed) Modified: stable/11/sys/dev/sfxge/sfxge.c ============================================================================== --- stable/11/sys/dev/sfxge/sfxge.c Tue Dec 25 07:38:46 2018 (r342454) +++ stable/11/sys/dev/sfxge/sfxge.c Tue Dec 25 07:39:34 2018 (r342455) @@ -760,6 +760,11 @@ sfxge_create(struct sfxge_softc *sc) } sc->rxq_entries = sfxge_rx_ring_entries; + if (efx_nic_cfg_get(enp)->enc_features & EFX_FEATURE_TXQ_CKSUM_OP_DESC) + sc->txq_dynamic_cksum_toggle_supported = B_TRUE; + else + sc->txq_dynamic_cksum_toggle_supported = B_FALSE; + if (!ISP2(sfxge_tx_ring_entries) || (sfxge_tx_ring_entries < EFX_TXQ_MINNDESCS) || (sfxge_tx_ring_entries > efx_nic_cfg_get(enp)->enc_txq_max_ndescs)) { Modified: stable/11/sys/dev/sfxge/sfxge.h ============================================================================== --- stable/11/sys/dev/sfxge/sfxge.h Tue Dec 25 07:38:46 2018 (r342454) +++ stable/11/sys/dev/sfxge/sfxge.h Tue Dec 25 07:39:34 2018 (r342455) @@ -290,6 +290,8 @@ struct sfxge_softc { efx_nic_t *enp; efsys_lock_t enp_lock; + boolean_t txq_dynamic_cksum_toggle_supported; + unsigned int rxq_entries; unsigned int txq_entries; Modified: stable/11/sys/dev/sfxge/sfxge_ev.c ============================================================================== --- stable/11/sys/dev/sfxge/sfxge_ev.c Tue Dec 25 07:38:46 2018 (r342454) +++ stable/11/sys/dev/sfxge/sfxge_ev.c Tue Dec 25 07:39:34 2018 (r342455) @@ -267,8 +267,11 @@ sfxge_get_txq_by_label(struct sfxge_evq *evq, enum sfx { unsigned int index; - KASSERT((evq->index == 0 && label < SFXGE_EVQ0_N_TXQ(evq->sc)) || - (label == SFXGE_TXQ_IP_TCP_UDP_CKSUM), ("unexpected txq label")); + KASSERT((evq->sc->txq_dynamic_cksum_toggle_supported) ? (label == 0) : + ((evq->index == 0 && label < SFXGE_TXQ_NTYPES) || + (label == SFXGE_TXQ_IP_TCP_UDP_CKSUM)), + ("unexpected txq label")); + index = (evq->index == 0) ? label : (evq->index - 1 + SFXGE_EVQ0_N_TXQ(evq->sc)); return (evq->sc->txq[index]); Modified: stable/11/sys/dev/sfxge/sfxge_tx.c ============================================================================== --- stable/11/sys/dev/sfxge/sfxge_tx.c Tue Dec 25 07:38:46 2018 (r342454) +++ stable/11/sys/dev/sfxge/sfxge_tx.c Tue Dec 25 07:39:34 2018 (r342455) @@ -33,7 +33,7 @@ /* Theory of operation: * - * Tx queues allocation and mapping + * Tx queues allocation and mapping on Siena * * One Tx queue with enabled checksum offload is allocated per Rx channel * (event queue). Also 2 Tx queues (one without checksum offload and one @@ -44,6 +44,17 @@ * if event queue index is 0, TxQ-index = TxQ-label * [0..SFXGE_TXQ_NTYPES) * else TxQ-index = SFXGE_TXQ_NTYPES + EvQ-index - 1 * See sfxge_get_txq_by_label() sfxge_ev.c + * + * Tx queue allocation and mapping on EF10 + * + * One Tx queue with enabled checksum offload is allocated per Rx + * channel (event queue). Checksum offload on all Tx queues is enabled or + * disabled dynamically by inserting option descriptors, so the additional + * queues used on Siena are not required. + * + * TxQ label is always set to zero on EF10 hardware. + * So, event queue to Tx queue mapping is simple: + * TxQ-index = EvQ-index */ #include @@ -137,38 +148,75 @@ static void sfxge_tx_qlist_post(struct sfxge_txq *txq) static void sfxge_tx_qunblock(struct sfxge_txq *txq); static int sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf, const bus_dma_segment_t *dma_seg, int n_dma_seg, - int vlan_tagged); + int n_extra_descs); +static inline void +sfxge_next_stmp(struct sfxge_txq *txq, struct sfxge_tx_mapping **pstmp) +{ + KASSERT((*pstmp)->flags == 0, ("stmp flags are not 0")); + if (__predict_false(*pstmp == + &txq->stmp[txq->ptr_mask])) + *pstmp = &txq->stmp[0]; + else + (*pstmp)++; +} + static int -sfxge_tx_maybe_insert_tag(struct sfxge_txq *txq, struct mbuf *mbuf) +sfxge_tx_maybe_toggle_cksum_offload(struct sfxge_txq *txq, struct mbuf *mbuf, + struct sfxge_tx_mapping **pstmp) { + uint16_t new_hw_cksum_flags; + efx_desc_t *desc; + + if (mbuf->m_pkthdr.csum_flags & + (CSUM_DELAY_DATA | CSUM_DELAY_DATA_IPV6 | CSUM_TSO)) { + /* + * We always set EFX_TXQ_CKSUM_IPV4 here because this + * configuration is the most useful, and this won't + * cause any trouble in case of IPv6 traffic anyway. + */ + new_hw_cksum_flags = EFX_TXQ_CKSUM_IPV4 | EFX_TXQ_CKSUM_TCPUDP; + } else if (mbuf->m_pkthdr.csum_flags & CSUM_DELAY_IP) { + new_hw_cksum_flags = EFX_TXQ_CKSUM_IPV4; + } else { + new_hw_cksum_flags = 0; + } + + if (new_hw_cksum_flags == txq->hw_cksum_flags) + return (0); + + desc = &txq->pend_desc[txq->n_pend_desc]; + efx_tx_qdesc_checksum_create(txq->common, new_hw_cksum_flags, desc); + txq->hw_cksum_flags = new_hw_cksum_flags; + txq->n_pend_desc++; + + sfxge_next_stmp(txq, pstmp); + + return (1); +} + +static int +sfxge_tx_maybe_insert_tag(struct sfxge_txq *txq, struct mbuf *mbuf, + struct sfxge_tx_mapping **pstmp) +{ uint16_t this_tag = ((mbuf->m_flags & M_VLANTAG) ? mbuf->m_pkthdr.ether_vtag : 0); + efx_desc_t *desc; if (this_tag == txq->hw_vlan_tci) return (0); - efx_tx_qdesc_vlantci_create(txq->common, - bswap16(this_tag), - &txq->pend_desc[0]); - txq->n_pend_desc = 1; + desc = &txq->pend_desc[txq->n_pend_desc]; + efx_tx_qdesc_vlantci_create(txq->common, bswap16(this_tag), desc); txq->hw_vlan_tci = this_tag; + txq->n_pend_desc++; + + sfxge_next_stmp(txq, pstmp); + return (1); } -static inline void -sfxge_next_stmp(struct sfxge_txq *txq, struct sfxge_tx_mapping **pstmp) -{ - KASSERT((*pstmp)->flags == 0, ("stmp flags are not 0")); - if (__predict_false(*pstmp == - &txq->stmp[txq->ptr_mask])) - *pstmp = &txq->stmp[0]; - else - (*pstmp)++; -} - - void sfxge_tx_qcomplete(struct sfxge_txq *txq, struct sfxge_evq *evq) { @@ -359,8 +407,9 @@ static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, int rc; int i; int eop; + uint16_t hw_cksum_flags_prev; uint16_t hw_vlan_tci_prev; - int vlan_tagged; + int n_extra_descs; KASSERT(!txq->blocked, ("txq->blocked")); @@ -411,14 +460,20 @@ static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, used_map = &stmp->map; + hw_cksum_flags_prev = txq->hw_cksum_flags; hw_vlan_tci_prev = txq->hw_vlan_tci; - vlan_tagged = sfxge_tx_maybe_insert_tag(txq, mbuf); - if (vlan_tagged) { - sfxge_next_stmp(txq, &stmp); - } + /* + * The order of option descriptors, which are used to leverage VLAN tag + * and checksum offloads, might be important. Changing checksum offload + * between VLAN option and packet descriptors probably does not work. + */ + n_extra_descs = sfxge_tx_maybe_toggle_cksum_offload(txq, mbuf, &stmp); + n_extra_descs += sfxge_tx_maybe_insert_tag(txq, mbuf, &stmp); + if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) { - rc = sfxge_tx_queue_tso(txq, mbuf, dma_seg, n_dma_seg, vlan_tagged); + rc = sfxge_tx_queue_tso(txq, mbuf, dma_seg, n_dma_seg, + n_extra_descs); if (rc < 0) goto reject_mapped; stmp = &txq->stmp[(rc - 1) & txq->ptr_mask]; @@ -429,7 +484,7 @@ static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, i = 0; for (;;) { - desc = &txq->pend_desc[i + vlan_tagged]; + desc = &txq->pend_desc[i + n_extra_descs]; eop = (i == n_dma_seg - 1); efx_tx_qdesc_dma_create(txq->common, dma_seg[i].ds_addr, @@ -441,7 +496,7 @@ static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, i++; sfxge_next_stmp(txq, &stmp); } - txq->n_pend_desc = n_dma_seg + vlan_tagged; + txq->n_pend_desc = n_dma_seg + n_extra_descs; } /* @@ -465,6 +520,7 @@ static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, reject_mapped: txq->hw_vlan_tci = hw_vlan_tci_prev; + txq->hw_cksum_flags = hw_cksum_flags_prev; bus_dmamap_unload(txq->packet_dma_tag, *used_map); reject: /* Drop the packet on the floor. */ @@ -838,8 +894,9 @@ sfxge_if_transmit(struct ifnet *ifp, struct mbuf *m) ("interface not up")); /* Pick the desired transmit queue. */ - if (m->m_pkthdr.csum_flags & - (CSUM_DELAY_DATA | CSUM_TCP_IPV6 | CSUM_UDP_IPV6 | CSUM_TSO)) { + if (sc->txq_dynamic_cksum_toggle_supported | + (m->m_pkthdr.csum_flags & + (CSUM_DELAY_DATA | CSUM_TCP_IPV6 | CSUM_UDP_IPV6 | CSUM_TSO))) { int index = 0; #ifdef RSS @@ -865,7 +922,9 @@ sfxge_if_transmit(struct ifnet *ifp, struct mbuf *m) if (m->m_pkthdr.csum_flags & CSUM_TSO) sfxge_parse_tx_packet(m); #endif - txq = sc->txq[SFXGE_TXQ_IP_TCP_UDP_CKSUM + index]; + index += (sc->txq_dynamic_cksum_toggle_supported == B_FALSE) ? + SFXGE_TXQ_IP_TCP_UDP_CKSUM : 0; + txq = sc->txq[index]; } else if (m->m_pkthdr.csum_flags & CSUM_DELAY_IP) { txq = sc->txq[SFXGE_TXQ_IP_CKSUM]; } else { @@ -1308,7 +1367,7 @@ static int tso_start_new_packet(struct sfxge_txq *txq, static int sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf, const bus_dma_segment_t *dma_seg, int n_dma_seg, - int vlan_tagged) + int n_extra_descs) { struct sfxge_tso_state tso; unsigned int id; @@ -1325,7 +1384,7 @@ sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf tso.in_len = dma_seg->ds_len - (tso.header_len - skipped); tso.dma_addr = dma_seg->ds_addr + (tso.header_len - skipped); - id = (txq->added + vlan_tagged) & txq->ptr_mask; + id = (txq->added + n_extra_descs) & txq->ptr_mask; if (__predict_false(tso_start_new_packet(txq, &tso, &id))) return (-1); @@ -1489,6 +1548,8 @@ sfxge_tx_qstop(struct sfxge_softc *sc, unsigned int in efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id, EFX_TXQ_NBUFS(sc->txq_entries)); + txq->hw_cksum_flags = 0; + SFXGE_EVQ_UNLOCK(evq); SFXGE_TXQ_UNLOCK(txq); } @@ -1510,6 +1571,10 @@ sfxge_tx_max_pkt_desc(const struct sfxge_softc *sc, en unsigned int fa_tso_v1_max_descs = 0; unsigned int fa_tso_v2_max_descs = 0; + /* Checksum offload Tx option descriptor may be required */ + if (sc->txq_dynamic_cksum_toggle_supported) + max_descs++; + /* VLAN tagging Tx option descriptor may be required */ if (efx_nic_cfg_get(sc->enp)->enc_hw_tx_insert_vlan_enabled) max_descs++; @@ -1554,6 +1619,7 @@ sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int i efsys_mem_t *esmp; uint16_t flags; unsigned int tso_fw_assisted; + unsigned int label; struct sfxge_evq *evq; unsigned int desc_index; int rc; @@ -1595,8 +1661,10 @@ sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int i break; } + label = (sc->txq_dynamic_cksum_toggle_supported) ? 0 : txq->type; + /* Create the common code transmit queue. */ - if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp, + if ((rc = efx_tx_qcreate(sc->enp, index, label, esmp, sc->txq_entries, txq->buf_base_id, flags, evq->common, &txq->common, &desc_index)) != 0) { /* Retry if no FATSOv2 resources, otherwise fail */ @@ -1606,7 +1674,7 @@ sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int i /* Looks like all FATSOv2 contexts are used */ flags &= ~EFX_TXQ_FATSOV2; tso_fw_assisted &= ~SFXGE_FATSOV2; - if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp, + if ((rc = efx_tx_qcreate(sc->enp, index, label, esmp, sc->txq_entries, txq->buf_base_id, flags, evq->common, &txq->common, &desc_index)) != 0) goto fail; @@ -1629,6 +1697,9 @@ sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int i txq->hw_vlan_tci = 0; + txq->hw_cksum_flags = flags & + (EFX_TXQ_CKSUM_IPV4 | EFX_TXQ_CKSUM_TCPUDP); + SFXGE_TXQ_UNLOCK(txq); return (0); @@ -1990,13 +2061,15 @@ sfxge_tx_init(struct sfxge_softc *sc) } /* Initialize the transmit queues */ - if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM, - SFXGE_TXQ_NON_CKSUM, 0)) != 0) - goto fail; + if (sc->txq_dynamic_cksum_toggle_supported == B_FALSE) { + if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM, + SFXGE_TXQ_NON_CKSUM, 0)) != 0) + goto fail; - if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM, - SFXGE_TXQ_IP_CKSUM, 0)) != 0) - goto fail2; + if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM, + SFXGE_TXQ_IP_CKSUM, 0)) != 0) + goto fail2; + } for (index = 0; index < sc->txq_count - SFXGE_EVQ0_N_TXQ(sc) + 1; Modified: stable/11/sys/dev/sfxge/sfxge_tx.h ============================================================================== --- stable/11/sys/dev/sfxge/sfxge_tx.h Tue Dec 25 07:38:46 2018 (r342454) +++ stable/11/sys/dev/sfxge/sfxge_tx.h Tue Dec 25 07:39:34 2018 (r342455) @@ -137,7 +137,9 @@ enum sfxge_txq_type { SFXGE_TXQ_NTYPES }; -#define SFXGE_EVQ0_N_TXQ(_sc) SFXGE_TXQ_NTYPES +#define SFXGE_EVQ0_N_TXQ(_sc) \ + ((_sc)->txq_dynamic_cksum_toggle_supported ? \ + 1 : SFXGE_TXQ_NTYPES) #define SFXGE_TXQ_UNBLOCK_LEVEL(_entries) (EFX_TXQ_LIMIT(_entries) / 4) @@ -203,6 +205,9 @@ struct sfxge_txq { unsigned int n_pend_desc; unsigned int added; unsigned int reaped; + + /* The last (or constant) set of HW offloads requested on the queue */ + uint16_t hw_cksum_flags; /* The last VLAN TCI seen on the queue if FW-assisted tagging is used */